MBZUAI's PAN 2 World Model: 29B Params, Fully Open with Weights, Inference and Training Code

ZhitingHu · x · 2026-11-20

The Institute of Foundation Models (IFM) at MBZUAI released PAN 2, a general-purpose world model under Apache 2.0 — shipping weights, inference code and training code, with the note that "open weights are the bare minimum of open." Capability: given a current observation (optional image + text) and text-described actions, PAN 2 simulates the outcome as up to 20 seconds of video, enabling simulative reasoning — an agent can imagine the outcomes of candidate actions before acting.

Architecture (GLP, from Critique of World Models): an LLM handles conceptual reasoning, writing the next world state as a structured, time-grounded description; a 29B MMDiT diffusion model trained from scratch on hundreds of millions of annotated videos fills in perceptual detail, with Wan 2.2 VAE decoding latents into video.

Original post →

More from Multimodal

Multimodal channel →